WP6: Computational models
Scientific background
Multimodal integration and disease progression modeling
Prof. Dr Yasser Iturria-Medina holds a Canada Research Chair and heads the Neuroinformatics for Personalized Medicine Laboratory at the Montreal Neurological Institute, McGill University. In the last 20 years, he has focused on defining and implementing multiscale and multifactorial brain-body computational models to characterize neurodegenerative diseases and identify effective personalized interventions. He has developed and validated multiple integrative molecular (genomics, epigenomics,
single-cell and bulk transcriptomics, proteomics, metabolomics, and lipidomics from brain and blood), neuroimaging (MRI, fMRI, DWI, PET), and clinical computational tools to understand the complex causal interactions between aging, neurodegeneration and various therapeutic conditions. These personalized models have allowed to characterize disease progression and heterogeneity in complex disorders like Alzheimer’s and PD, and are shared by Prof. Iturria-Medina via a free cross-platform user-friendly multitool software (NeuroPM-box) for advanced computational integration of post-mortem and in-vivo molecular, histopathological, multimodal neuroimaging, and therapeutic data in neurology.
Working plan and methodology
Modeling
This WP is the culminating aspect of the project. We will apply two complementary types of multifactorial computational disease models to extensively characterize PD progression and heterogeneity in the early stages, benefiting from a large, uniquely rich, comprehensive dataset acquired in WP1, 3, 4 and 5. First, we will use all the generated imaging-derived metrics and advanced machine learning-based trajectory
inference techniques47 to characterize temporal disease progression and heterogeneity in PD. This will allow us to identify distinctive biologically-defined PD subtrajectories aligned with clinical deterioration, while reordering participants in terms of PD stage and putative subtypes. Our unsupervised machine learning methods can integrate large number of biological markers (e.g., voxel-wise, cortical layer-wise and region-wise MRI-derived tissue microstructure metrics), allowing us to also identify the best markers capturing PD progression and heterogeneity, and how they associate with other relevant factors (e.g., genetics, sleep, behavior). Specifically, we will apply the multimodal contrastive subtrajectories inference algorithm (mcTI), which was developed by Prof. Itturia-Medina’s group and extensively tested/validated in the context of large-scale multi-omics molecular unification in Alzheimer’s disease47. Using recent advances in artificial intelligence (AI) to explore and visualize high dimensional data21, the mcTI will be used to detect PD-associated multimodal neuroimaging patterns, anchored in clinical normality (NC) relative to advanced PD. All neuroimaging-derived markers will be first adjusted for potentially confounding covariates (e.g., age, sex, education, site, MRI protocol). Similar to our previous work, which was focused on four different molecular omics in brain and blood, here we will assume that the position of each subject in the aggregated biological space would predict individual severity of PD and proximity
to distinctive disease subtrajectories. For the first time, a unified multi-level PD Progression Score (PD-PS) will be calculated for each subject, ranging from 0 (indicating NC) to 1 (representing late-stage PD). Furthermore, each participant will undergo PD subtyping based on the highest likelihood of aligning with
distinctive disease subtrajectories in the neuroimaging-based disease space. Once the distinct neuroimaging-based subtrajectories of PD progression and heterogeneity are identified, ANOVA tests will then be used to compare these subtrajectories in terms of genetic, behavioral, clinical, and demographic
characteristics as well as in terms of PD-PS. A notable attribute of the mcTI approach is that it allows estimating the specific contribution of each neuroimaging marker to the identified disease “timeline” and the obtained personalized disease index and subtype. Simply put, this technique overcomes the traditional AI “black-box” limitation, and allow for discovering and interpreting neuroimaging metrics and brain regions strongly associated with PD evolution and heterogeneity.
Next, we will use mechanistic causal models of disease cerebral spreading, integrating longitudinal qMRI,
volumetric MRI, DWI, and fMRI to characterize how different biological factors/markers directly interact and how pathological alterations on these factors propagate through the brain via the individual connectomes (i.e., multifactorial spreading of the disease). Additionally, integration with whole-brain neurotransmitter receptor maps will allow us to identify genes and receptors associated with intra-brain pathological interactions and the across-region disease spreading as described. First, the multifactorial causal model (MCM) will be applied to the multimodal longitudinal data. Based on dynamic system analysis, MCM allows the data-driven exploration of multiple interacting cerebral and clinical factors to model brain (dis)organization and therapeutic interventions. Specifically, this computational framework will be used at the individual level to: i) quantify pairwise direct interactions between all neuroimaging-derived biological factors (iron content, tissue atrophy, microstructural properties and connectivity), ii) identify the most-likely brain region and neuroimaging-derived biological factor initiating the pathologic cascade leading to PD, and iii) characterize the network-mediated intra-brain propagation/spreading of all the neuroimaging-derived biological factors considered. All individuals will be characterized by a set of biologically-interpretable model parameters, which will allow direct comparison between them and subsequent association with clinical severity, genetic profile, behavior, and other relevant factors (e.g.,
sleep patterns). Next, we will use extended MCM versions to investigate the role of genes and neurotransmitter receptors in the multi-faceted neurodegeneration processes of PD. By combining whole-brain spatial maps of 998 genes and 15 neurotransmitter receptors with the longitudinal neuroimaging data from the analyzed cohort, personalized generative models (gene expression- and receptor-enriched multifactorial causal models, ge/re-MCM) will be applied to: i) test improvement in the explanation of neuroimaging alterations compared to models without receptor data, ii) quantify stable gene- and receptor-mediated mechanisms involved in the pairwise interactions
between the neuroimaging-derived biological factors (i.e., estimating each gene and neuroreceptor’s impact on interactions among iron content, tissue atrophy, microstructural properties and connectivity), iii) identify distinct axes of gene- and receptor-pathology alterations in terms of their association with clinical symptoms, and iv) map brain regions where specific genes’ expression and receptor densities significantly influence different types of neuropathology, highlighting the regional heterogeneity of genetic and receptor involvement in PD neurodegeneration.
Team - Montreal Neurological Institute, McGill University
